This paper describes our system for SemEval-2021 Task 5 on Toxic Spans Detection. We developed ensemble models using BERT-based neural architectures and post-processing to combine tokens into spans. We evaluated several pre-trained language models using various ensemble techniques for toxic span identification and achieved sizable improvements over our baseline fine-tuned BERT models. Finally, our system obtained a F1-score of 67.55% on test data.
@article{arxiv.2104.04739,
title = {MIPT-NSU-UTMN at SemEval-2021 Task 5: Ensembling Learning with Pre-trained Language Models for Toxic Spans Detection},
author = {Mikhail Kotyushev and Anna Glazkova and Dmitry Morozov},
journal= {arXiv preprint arXiv:2104.04739},
year = {2021}
}
Comments
Accepted at SemEval-2021 Workshop, ACL-IJCNLP 2021